PFN releases PLaMo 3.0 Fin, a finance-focused small language model with open weights that runs from a single GPU in closed environments
Synopsis
Preferred Networks built PLaMo 3.0 Fin on PLaMo 3 NICT, a pretrained model from joint research with NICT, by further training on Japanese finance data, and released it as an open-weight finance-focused small language model. Aimed at financial institutions' operations, it reasons before answering with the reasoning trace in Japanese and can call external tools such as web search, calculation and internal databases. Its small size lets it run from a single GPU in closed cloud or on-premises environments, positioned as the core model of AI agents handling tasks such as inquiry classification, document extraction and alert analysis.
Interpretation
Financial institutions handle large volumes of routine processing, monitoring, client management and inquiry response, and many of these tasks demand fast turnaround; running all of them through a large language model raises usage costs and may not pay off. Bringing generative AI into financial practice previously required, beyond financial expertise, handling Japanese business conventions and regulations, institution-specific terminology and workflows, and running efficiently in closed environments that handle highly confidential customer and transaction data. PLaMo 3.0 Fin is based on PLaMo 3 NICT, the pretrained model from PFN's joint research with NICT, further trained by PFN on Japanese finance data and released as open weights on Hugging Face under the PLaMo Community License.
The model is intended as the core of AI agents that take on part of a financial institution's business processes, and it reasons before answering, with the reasoning trace output in Japanese so the basis of an answer can be checked. It also adds the ability to call external tools for web search of the latest information, calculation and reference to internal databases. Because it is a small model that runs fast, it can be deployed in closed cloud and on-premises environments starting from a single GPU; the released open-weight model is named plamo-3-8b-fin-instruct.
In post-training for PLaMo 3.0 Fin, instruction-following training for reasoning and tool use was carried out on MN-Core 2, PFN's in-house AI semiconductor. For more developers to use an in-house AI semiconductor, the specialized knowledge and development burden of porting programs must be reduced without giving up hardware performance. PFN's new porting approach, centered on a coding agent, let developers without deep MN-Core 2 expertise complete porting of the instruction-following training code, including support for multi-board distributed training, within one week; training with the ported code reached practical training speed comparable to a GPU training environment.
Perspective
The public scope is the open weights of the finance-focused small language model, the free plan of the API and Chat (beta), and a service for building institution-specific models, offered since June 2025. The material addresses financial institutions and developers considering deployment of generative AI in closed or on-premises environments, and gives no quantitative evaluation results on financial tasks.
The API and Chat (beta) have usage caps, their content may change or end without notice, and no service level such as availability is guaranteed; the demo output is only an example, and the same input may yield different results.